re: Issue https://github.com/Unidata/netcdf-c/issues/2685
re: PR https://github.com/Unidata/netcdf-c/pull/2179
As noted in PR https://github.com/Unidata/netcdf-c/pull/2179,
the old code did not allow for reclaiming instances of types,
nor for properly copying them. That PR provided new functions
capable of reclaiming/copying instances of arbitrary types.
However, as noted by Issue https://github.com/Unidata/netcdf-c/issues/2685, using these
most general functions resulted in a significant performance
degradation, even for common cases.
This PR attempts to mitigate the cost of using the general
reclaim/copy functions in two ways.
First, the previous functions operating at the top level by
using ncid and typeid arguments. These functions were augmented
with equivalent versions that used the netcdf-c library internal
data structures to allow direct access to needed information.
These new functions are used internally to the library.
The second mitigation involves optimizing the internal functions
by providing early tests for common cases. This avoids
unnecessary recursive function calls.
The overall result is a significant improvement in speed by a
factor of roughly twenty -- your mileage may vary. These
optimized functions are still not as fast as the original (more
limited) functions, but they are getting close. Additional optimizations are
possible. But the cost is a significant "uglification" of the
code that I deemed a step too far, at least for now.
## Misc. Changes
1. Added a test case to check the proper reclamation/copy of complex types.
2. Found and fixed some places where nc_reclaim/copy should have been used.
3. Replaced, in the netcdf-c library, (almost all) occurrences of nc_reclaim_copy with calls to NC_reclaim/copy. This plus the optimizations is the primary speed-up mechanism.
4. In DAP4, the metadata is held in a substrate in-memory file; this required some changes so that the reclaim/copy code accessed that substrate dispatcher rather than the DAP4 dispatcher.
5. Re-factored and isolated the code that computes if a type is (transitively) variable-sized or not.
6. Clean up the reclamation code in ncgen; adding the use of nc_reclaim exposed some memory problems.
## Improvements to S3 Documentation
* Create a new document *quickstart_paths.md* that give a summary of the legal path formats used by netcdf-c. This includes both file paths and URL paths.
* Modify *nczarr.md* to remove most of the S3 related text.
* Move the S3 text from *nczarr.md* to a new document *cloud.md*.
* Add some S3-related text to the *byterange.md* document.
Hopefully, this will make it easier for users to find the information they want.
## Rebuild NCZarr Testing
In order to avoid problems with running make check in parallel, two changes were made:
1. The *nczarr_test* test system was rebuilt. Now, for each test.
any generated files are kept in a test-specific directory, isolated
from all other test executions.
2. Similarly, since the S3 test bucket is shared, any generated S3 objects
are isolated using a test-specific key path.
## Other S3 Related Changes
* Add code to ensure that files created on S3 are reclaimed at end of testing.
* Used the bash "trap" command to ensure S3 cleanup even if the test fails.
* Cleanup the S3 related configure.ac flag set since S3 is used in several places. So now one should use the option *--enable-s3* instead of *--enable-nczarr-s3*, although the latter is still kept as a deprecated alias for the former.
* Get some of the github actions yml to work with S3; required fixing various test scripts adding a secret to access the Unidata S3 bucket.
* Cleanup S3 portion of libnetcdf.settings.in and netcdf_meta.h.in and test_common.in.
* Merge partial S3 support into dhttp.c.
* Create an experimental s3 access library especially for use with Windows. It is enabled by using the options *--enable-s3-internal* (automake) or *-DENABLE_S3_INTERNAL=ON* (CMake). Also add a unit-test for it.
* Move some definitions from ncrc.h to ncs3sdk.h
## Other Changes
* Provide a default implementation of strlcpy and move this and similar defaults into *dmissing.c*.
re: https://github.com/Unidata/netcdf-c/issues/541
re: https://github.com/Unidata/netcdf-c/issues/1208
re: https://github.com/Unidata/netcdf-c/issues/2078
re: https://github.com/Unidata/netcdf-c/issues/2041
re: https://github.com/Unidata/netcdf-c/issues/2143
For a long time, there have been known problems with the
management of complex types containing VLENs. This also
involves the string type because it is stored as a VLEN of
chars.
This PR (mostly) fixes this problem. But note that it adds new
functions to netcdf.h (see below) and this may require bumping
the .so number. These new functions can be removed, if desired,
in favor of functions in netcdf_aux.h, but netcdf.h seems the
better place for them because they are intended as alternatives
to the nc_free_vlen and nc_free_string functions already in
netcdf.h.
The term complex type refers to any type that directly or
transitively references a VLEN type. So an array of VLENS, a
compound with a VLEN field, and so on.
In order to properly handle instances of these complex types, it
is necessary to have function that can recursively walk
instances of such types to perform various actions on them. The
term "deep" is also used to mean recursive.
At the moment, the two operations needed by the netcdf library are:
* free'ing an instance of the complex type
* copying an instance of the complex type.
The current library does only shallow free and shallow copy of
complex types. This means that only the top level is properly
free'd or copied, but deep internal blocks in the instance are
not touched.
Note that the term "vector" will be used to mean a contiguous (in
memory) sequence of instances of some type. Given an array with,
say, dimensions 2 X 3 X 4, this will be stored in memory as a
vector of length 2*3*4=24 instances.
The use cases are primarily these.
## nc_get_vars
Suppose one is reading a vector of instances using nc_get_vars
(or nc_get_vara or nc_get_var, etc.). These functions will
return the vector in the top-level memory provided. All
interior blocks (form nested VLEN or strings) will have been
dynamically allocated.
After using this vector of instances, it is necessary to free
(aka reclaim) the dynamically allocated memory, otherwise a
memory leak occurs. So, the recursive reclaim function is used
to walk the returned instance vector and do a deep reclaim of
the data.
Currently functions are defined in netcdf.h that are supposed to
handle this: nc_free_vlen(), nc_free_vlens(), and
nc_free_string(). Unfortunately, these functions only do a
shallow free, so deeply nested instances are not properly
handled by them.
Note that internally, the provided data is immediately written so
there is no need to copy it. But the caller may need to reclaim the
data it passed into the function.
## nc_put_att
Suppose one is writing a vector of instances as the data of an attribute
using, say, nc_put_att.
Internally, the incoming attribute data must be copied and stored
so that changes/reclamation of the input data will not affect
the attribute.
Again, the code inside the netcdf library does only shallow copying
rather than deep copy. As a result, one sees effects such as described
in Github Issue https://github.com/Unidata/netcdf-c/issues/2143.
Also, after defining the attribute, it may be necessary for the user
to free the data that was provided as input to nc_put_att().
## nc_get_att
Suppose one is reading a vector of instances as the data of an attribute
using, say, nc_get_att.
Internally, the existing attribute data must be copied and returned
to the caller, and the caller is responsible for reclaiming
the returned data.
Again, the code inside the netcdf library does only shallow copying
rather than deep copy. So this can lead to memory leaks and errors
because the deep data is shared between the library and the user.
# Solution
The solution is to build properly recursive reclaim and copy
functions and use those as needed.
These recursive functions are defined in libdispatch/dinstance.c
and their signatures are defined in include/netcdf.h.
For back compatibility, corresponding "ncaux_XXX" functions
are defined in include/netcdf_aux.h.
````
int nc_reclaim_data(int ncid, nc_type xtypeid, void* memory, size_t count);
int nc_reclaim_data_all(int ncid, nc_type xtypeid, void* memory, size_t count);
int nc_copy_data(int ncid, nc_type xtypeid, const void* memory, size_t count, void* copy);
int nc_copy_data_all(int ncid, nc_type xtypeid, const void* memory, size_t count, void** copyp);
````
There are two variants. The first two, nc_reclaim_data() and
nc_copy_data(), assume the top-level vector is managed by the
caller. For reclaim, this is so the user can use, for example, a
statically allocated vector. For copy, it assumes the user
provides the space into which the copy is stored.
The second two, nc_reclaim_data_all() and
nc_copy_data_all(), allows the functions to manage the
top-level. So for nc_reclaim_data_all, the top level is
assumed to be dynamically allocated and will be free'd by
nc_reclaim_data_all(). The nc_copy_data_all() function
will allocate the top level and return a pointer to it to the
user. The user can later pass that pointer to
nc_reclaim_data_all() to reclaim the instance(s).
# Internal Changes
The netcdf-c library internals are changed to use the proper
reclaim and copy functions. It turns out that the places where
these functions are needed is quite pervasive in the netcdf-c
library code. Using these functions also allows some
simplification of the code since the stdata and vldata fields of
NC_ATT_INFO are no longer needed. Currently this is commented
out using the SEPDATA \#define macro. When any bugs are largely
fixed, all this code will be removed.
# Known Bugs
1. There is still one known failure that has not been solved.
All the failures revolve around some variant of this .cdl file.
The proximate cause of failure is the use of a VLEN FillValue.
````
netcdf x {
types:
float(*) row_of_floats ;
dimensions:
m = 5 ;
variables:
row_of_floats ragged_array(m) ;
row_of_floats ragged_array:_FillValue = {-999} ;
data:
ragged_array = {10, 11, 12, 13, 14}, {20, 21, 22, 23}, {30, 31, 32},
{40, 41}, _ ;
}
````
When a solution is found, I will either add it to this PR or post a new PR.
# Related Changes
* Mark nc_free_vlen(s) as deprecated in favor of ncaux_reclaim_data.
* Remove the --enable-unfixed-memory-leaks option.
* Remove the NC_VLENS_NOTEST code that suppresses some vlen tests.
* Document this change in docs/internal.md
* Disable the tst_vlen_data test in ncdump/tst_nccopy4.sh.
* Mark types as fixed size or not (transitively) to optimize the reclaim
and copy functions.
# Misc. Changes
* Make Doxygen process libdispatch/daux.c
* Make sure the NC_ATT_INFO_T.container field is set.
Filter support has three goals:
1. Use the existing HDF5 filter implementations,
2. Allow filter metadata to be stored in the NumCodecs metadata format used by Zarr,
3. Allow filters to be used even when HDF5 is disabled
Detailed usage directions are define in docs/filters.md.
For now, the existing filter API is left in place. So filters
are defined using ''nc_def_var_filter'' using the HDF5 style
where the id and parameters are unsigned integers.
This is a big change since filters affect many parts of the code.
In the following, the terms "compressor" and "filter" and "codec" are generally
used synonomously.
### Filter-Related Changes:
* In order to support dynamic loading of shared filter libraries, a new library was added in the libncpoco directory; it helps to isolate dynamic loading across multiple platforms.
* Provide a json parsing library for use by plugins; this is created by merging libdispatch/ncjson.c with include/ncjson.h.
* Add a new _Codecs attribute to allow clients to see what codecs are being used; let ncdump -s print it out.
* Provide special headers to help support compilation of HDF5 filters when HDF5 is not enabled: netcdf_filter_hdf5_build.h and netcdf_filter_build.h.
* Add a number of new test to test the new nczarr filters.
* Let ncgen parse _Codecs attribute, although it is ignored.
### Plugin directory changes:
* Add support for the Blosc compressor; this is essential because it is the most common compressor used in Zarr datasets. This also necessitated adding a CMake FindBlosc.cmake file
* Add NCZarr support for the big-four filters provided by HDF5: shuffle, fletcher32, deflate (zlib), and szip
* Add a Codec defaulter (see docs/filters.md) for the big four filters.
* Make plugins work with windows by properly adding __declspec declaration.
### Misc. Non-Filter Changes
* Replace most uses of USE_NETCDF4 (deprecated) with USE_HDF5.
* Improve support for caching
* More fixes for path conversion code
* Fix misc. memory leaks
* Add new utility -- ncdump/ncpathcvt -- that does more or less the same thing as cygpath.
* Add a number of new test to test the non-filter fixes.
* Update the parsers
* Convert most instances of '#ifdef _MSC_VER' to '#ifdef _WIN32'
The XArray implementation that uses Zarr for storage
provides a mechanism to simulate named dimensions.
It does this by adding a per-variable attribute called
_ARRAY_DIMENSIONS. This attribute contains a list of names
to be matched against the shape values of the variable.
In effect a named dimension is created with the name
_ARRAY_DIMENSIONS(i) and length shape(i) for all i
in range 0..rank(variable).
Both read and write support is provided.
This XArray support is only invoked if the mode value
of "xarray" is defined. So for example, as in this URL.
````
https://s3.us-west-1.amazonaws.com/bucket/dataset#mode=nczarr,xarray,s3
````
Note that the "xarray" mode flag also implies mode flag "zarr", so the above
is equivalent to this URL.
````
https://s3.us-west-1.amazonaws.com/bucket/dataset#mode=nczarr,zarr,xarray,s3
````
The primary change to implement this was to unify the handling
of dimension references in libnczarr/zsync.
A test for this and other pure-zarr features was added as
nczarr_test/run_purezarr.sh
Other changes:
* Make sure distcheck leaves no files around.
* Change the special attribute flag DIMSCALEFLAG to HIDDENATTRFLAG
to support the xarray attribute.
* Annotate the zmap implementations with feature flags such as
WRITEONCE (for zip files).
Primary changes:
* Add an improved cache system to speed up performance.
* Fix NCZarr to properly handle scalar variables.
Misc. Related Changes:
* Added unit tests for extendible hash and for the generic cache.
* Add config parameter to set size of the NCZarr cache.
* Add initial performance tests but leave them unused.
* Add CRC64 support.
* Move location of ncdumpchunks utility from /ncgen to /ncdump.
* Refactor auth support.
Misc. Unrelated Changes:
* More cleanup of the S3 support
* Add support for S3 authentication in .rc files: HTTP.S3.ACCESSID and HTTP.S3.SECRETKEY.
* Remove the hashkey from the struct OBJHDR since it is never used.
cloud using a variant of the Zarr protocol and storage
format. This enhancement is generically referred to as "NCZarr".
The data model supported by NCZarr is netcdf-4 minus the user-defined
types and the String type. In this sense it is similar to the CDF-5
data model.
More detailed information about enabling and using NCZarr is
described in the document NUG/nczarr.md and in a
[Unidata Developer's blog entry](https://www.unidata.ucar.edu/blogs/developer/en/entry/overview-of-zarr-support-in).
WARNING: this code has had limited testing, so do use this version
for production work. Also, performance improvements are ongoing.
Note especially the following platform matrix of successful tests:
Platform | Build System | S3 support
------------------------------------
Linux+gcc | Automake | yes
Linux+gcc | CMake | yes
Visual Studio | CMake | no
Additionally, and as a consequence of the addition of NCZarr,
major changes have been made to the Filter API. NOTE: NCZarr
does not yet support filters, but these changes are enablers for
that support in the future. Note that it is possible
(probable?) that there will be some accidental reversions if the
changes here did not correctly mimic the existing filter testing.
In any case, previously filter ids and parameters were of type
unsigned int. In order to support the more general zarr filter
model, this was all converted to char*. The old HDF5-specific,
unsigned int operations are still supported but they are
wrappers around the new, char* based nc_filterx_XXX functions.
This entailed at least the following changes:
1. Added the files libdispatch/dfilterx.c and include/ncfilter.h
2. Some filterx utilities have been moved to libdispatch/daux.c
3. A new entry, "filter_actions" was added to the NCDispatch table
and the version bumped.
4. An overly complex set of structs was created to support funnelling
all of the filterx operations thru a single dispatch
"filter_actions" entry.
5. Move common code to from libhdf5 to libsrc4 so that it is accessible
to nczarr.
Changes directly related to Zarr:
1. Modified CMakeList.txt and configure.ac to support both C and C++
-- this is in support of S3 support via the awd-sdk libraries.
2. Define a size64_t type to support nczarr.
3. More reworking of libdispatch/dinfermodel.c to
support zarr and to regularize the structure of the fragments
section of a URL.
Changes not directly related to Zarr:
1. Make client-side filter registration be conditional, with default off.
2. Hack include/nc4internal.h to make some flags added by Ed be unique:
e.g. NC_CREAT, NC_INDEF, etc.
3. cleanup include/nchttp.h and libdispatch/dhttp.c.
4. Misc. changes to support compiling under Visual Studio including:
* Better testing under windows for dirent.h and opendir and closedir.
5. Misc. changes to the oc2 code to support various libcurl CURLOPT flags
and to centralize error reporting.
6. By default, suppress the vlen tests that have unfixed memory leaks; add option to enable them.
7. Make part of the nc_test/test_byterange.sh test be contingent on remotetest.unidata.ucar.edu being accessible.
Changes Left TO-DO:
1. fix provenance code, it is too HDF5 specific.
nc4internal.c contains code to free the format_XXX_info
fields. Since these are format specific, this code
was moved to the dispatch code (libhdf5 and libhdf4
in the current case).
Additionally, there are some fields in nc4internal.h (e.g.
dimscale fields) that are specific to HDF5 and have been moved
to the corresponding HDF5 data structures and code.
Misc. other changes:
1. NC_VAR_INFO_T->hdf5_name renamed to alt_name to avoid
implying it is necessarily HDF5 specific.
2. prefix NC_FILE_INFO_T with an instance of NC_OBJ for consistency.
this also requires wrapping move_in_NCList() to keep
hdr.id consistent.
re: https://github.com/Unidata/netcdf-c/issues/1352
When nc4info.c encounters an _NCProperties attribute
with a version number it does not recognize, it does not
show it correctly.
Solution chosen is to arrange so that accessing the attribute
returns the raw value of the Attribute from the file. This way,
even if the version is unrecognized, it will return something
usable.
The changes were primarily to never attempt to parse the value
of _NCProperties until actually required. Which since they
are currently not used means that parsing never occurs.
Also modified ncdump/tst_fileinfo.sh to include some extra testing
I tested the original failure by changing the value of NCPROPS to 3.
However, there is no way to test this at build time.
Misc. Changes
* Inlined the provenance info in the NC_FILE_INFO_T structure
* Centralized stuff from elsewhere into include/nc_provenance.h
Misc. Unrelated Changes
* Removed/turned off some misc debug output left on by accident
* Fix CPPFLAGS name error in libhdf5/Makefile.am
This is a follow up to PR https://github.com/Unidata/netcdf-c/pull/1173
Sorry that it is so big, but leak suppression can be complex.
This PR fixes all remaining memory leaks -- as determined by
-fsanitize=address, and with the exceptions noted below.
Unfortunately. there remains a significant leak that I cannot
solve. It involves vlens, and it is unclear if the leak is
occurring in the netcdf-c library or the HDF5 library.
I have added a check_PROGRAM to the ncdump directory to show the
problem. The program is called tst_vlen_demo.c To exercise it,
build the netcdf library with -fsanitize=address enabled. Then
go into ncdump and do a "make clean check". This should build
tst_vlen_demo without actually executing it. Then do the
command "./tst_vlen_demo" to see the output of the memory
checker. Note the the lost malloc is deep in the HDF5 library
(in H5Tvlen.c).
I am temporarily working around this error in the following way.
1. I modified several test scripts to not execute known vlen tests
that fail as described above.
2. Added an environment variable called NC_VLEN_NOTEST.
If set, then those specific tests are suppressed.
This should mean that the --disable-utilities option to
./configure should not need to be set to get a memory leak clean
build. This should allow for detection of any new leaks.
Note: I used an environment variable rather than a ./configure
option to control the vlen tests. This is because it is
temporary (I hope) and because it is a bit tricky for shell
scripts to access ./configure options.
Finally, as before, this only been tested with netcdf-4 and hdf5 support.
stored in the _NCProperties attribute to allow two things:
1. capture of additional library dependencies (over and above
hdf5)
2. Recognition of non-netcdf libraries that create netcdf-4 format
files.
To this end, the _NCProperties format has been extended to be
and arbitrary set of key=value pairs separated by commas.
This new format has version = 2, and uses commas as the pair separator.
Thus the general form is:
_NCProperties = "version=2,key1=value,key2=value2..." ;
This new version is accompanied by a new ./configure option of the form
--with-ncproperties="key1=value1,key2=value2..."
that specifies pairs to add to the _NCProperties attribute for all
files created with that netcdf library.
At this point, what is missing is some programmatic way to
specify either all the pairs or additional pairs
to the _NCProperties attribute. Not sure of the best way
to do this.
Builders using non-netcdf libraries can specify
whatever they want in the key value pairs (as long
as the version=2 is specified first).
By convention, the primary library is expected to be the
the first pair after the leading version=2 pair, but this
is convention only and is neither required nor enforced.
Related changes:
1. Fixed the tests that check _NCProperties to properly operate with version=2.
2. When reading a version 1 _NCProperties attribute, convert it to look
like a version 2 attribute.
2. Added some version 2 tests to ncdump/tst_fileinfo.c and
ncdump/tst_fileinfo.sh
Misc Changes:
1. Fix minor problem in ncdap_test/testurl.sh where a parameter to
buildurl needed to be quoted.
2. Minor fix to ncgen to swap switches -H and -h to be consistent
with other utilities.
3. Document the -M flag in nccopy usage() and the nccopy man page.
4. Modify a test case to use the nccopy -M flag.
